Evaluation
Aggregates, comparisons and benchmark reports.
VISUALIZATION
Inspect the evidence behind a metric.
EvalScope reports connect aggregate results with sample reviews, agent details and serving charts. Run locally; no hosted customer data is required.
FROM OUTPUTS TO INSIGHTS
One local dashboard, one evidence trail.
Same reports · better stays with you
1
Outputs directory
outputs/
2
Scan runs
Discover evaluations.
3
Compare
Contrast results.
4
Inspect samples
Open reviews.
5
Share report
HTML / JSON
FOUR EVIDENCE VIEWS
Agent
Task details, model requests and tool event evidence.
Performance
Load curves, latency, TTFT and throughput charts.
Run artifacts
Configuration, raw outputs and shareable HTML reports.
Four views, one evidence trail.
Real Dashboard screenshots
Aggregate → sample → artifact.
Use the report to locate the sample, then its prediction and review. The same output folder preserves configs and raw artifacts for a local audit.
pip install 'evalscope[service]'
evalscope service --outputs ./outputsLOCAL EVIDENCE, NOT A DETACHED CHART
Keep the chart tied to the run that produced it.
A report is useful when you can move from a number to its samples, config and raw artifacts without losing the thread.
See the evaluation path01reports/Aggregate decisions
02reviews/Sample evidence
03configs/Runnable context
04outputs/Raw record